inference-relay 1.4.0

package/dist/index.js

@@ -0,0 +1,24 @@
+import {execSync}from'child_process';import {mkdirSync,writeFileSync,existsSync,readFileSync}from'fs';import {homedir}from'os';import {join,dirname}from'path';import {createHash,randomUUID}from'crypto';import {compileFunction}from'vm';import {createRequire}from'module';import {EventEmitter}from'events';/*! inference-relay (c) 2026 L2B II LLC. Patent pending. Proprietary. See LICENSE. */
+var je=Object.defineProperty;var Je=(n=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(n,{get:(e,t)=>(typeof require<"u"?require:e)[t]}):n)(function(n){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+n+'" is not supported')});var ze=(n,e)=>()=>(n&&(e=n(n=0)),e);var Xe=(n,e)=>{for(var t in e)je(n,t,{get:e[t],enumerable:true});};var Me={};Xe(Me,{detectCLI:()=>H});async function H(){let n=null;try{n=execSync("which claude",{encoding:"utf-8",timeout:5e3}).trim();}catch{return {available:false,authenticated:false}}if(!n)return {available:false,authenticated:false};let e=null;try{let r=execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!at.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var at,ee=ze(()=>{at=["2.1.87"];});var U=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};function L(n,e){let t=new Map,r=null,s=null,i=new Promise((u,g)=>{r=u,s=g;});i.catch(()=>{});let a=[],l=null,o="",d="";function c(u,...g){let m=t.get(u);m&&m.forEach(y=>y(...g));}async function*f(){try{for await(let u of n)if(yield u,u.type==="content_block_start")a[u.index]=u.content_block;else if(u.type==="content_block_delta"){let g=a[u.index];g&&u.delta?.type==="text_delta"&&(g.text=(g.text||"")+u.delta.text,c("text",u.delta.text));}else if(u.type==="message_start")o=u.message?.model||"",l=u.message?.usage;else if(u.type==="message_delta")d=u.delta?.stop_reason||"",u.usage&&(l={...l,...u.usage});else if(u.type==="message_stop"){let g={role:"assistant",model:o,content:a.filter(Boolean),usage:l,stop_reason:d};c("message",g),r&&(r(g),r=null),e&&e(g);}if(r){let u={role:"assistant",model:o,content:a.filter(Boolean),usage:l,stop_reason:d||"end_turn"};c("message",u),r(u),r=null,e&&e(u);}}catch(u){throw c("error",u),s&&(s(u),s=null),u}}let h=f(),p=false,E={[Symbol.asyncIterator](){if(p)throw new Error("[inference-relay] RelayStream can only be iterated once");return p=true,h},on(u,g){return t.has(u)||t.set(u,[]),t.get(u).push(g),E},finalMessage(){let u,g=new Promise((m,y)=>{u=setTimeout(()=>y(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([i,g]).finally(()=>clearTimeout(u))}};return E}var A=class{priority=100;async streamRaw(e,t){throw new Error(`[inference-relay] Provider '${this.type}' does not implement streamRaw(). Override either stream() or streamRaw() in the provider class.`)}async stream(e,t){return L(await this.streamRaw(e,t))}};var pe=`-----BEGIN PUBLIC KEY-----
+MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
+EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
+t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
+kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
+7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
+Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
+UQIDAQAB
+-----END PUBLIC KEY-----`,de={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},M=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(pe,"RS256"),{payload:i}=await t(e,s),a=JSON.parse(new TextDecoder().decode(i));return this.verificationFailures=0,this.cachedManifest=a,this.currentSecurityState="SEC_NOMINAL",{manifest:a,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:de,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:de,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},Qe=new M;function Pe(){Qe.clearCache();}function Ze(){return de}function et(n,e){return e.split(".").reduce((t,r)=>{if(!(r==="__proto__"||r==="constructor"||r==="prototype"))return t&&t[r]},n)}var it=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";function ye(n){return join(homedir(),".inference-relay","gateway",n)}function Ie(n){return join(ye(n),"bundle.js")}function Te(n){return join(ye(n),"meta.json")}var _=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},F=new Map,B=class{static async load(e,t,r=it){let s=`${t}:${e.slice(-8)}`;if(F.has(s))return F.get(s);try{let f=await this.tryLoadFromDisk(t);if(f)return F.set(s,f),f}catch{}let{bytes:i,sha256:a,signature:l,libraryVersion:o}=await this.fetchFromAuthority(e,t,r);if(createHash("sha256").update(i).digest("hex")!==a)throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(l,o,a);try{let f=ye(t);mkdirSync(f,{recursive:!0}),writeFileSync(Ie(t),i),writeFileSync(Te(t),JSON.stringify({sha256:a,signature:l,libraryVersion:o}));}catch{}let c=this.loadIntoVm(i.toString("utf-8"));return F.set(s,c),c}static async loadFromDisk(e){if(!existsSync(e))throw new _(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Ie(e),r=Te(e);if(!existsSync(t)||!existsSync(r))return null;let s=readFileSync(t),i=JSON.parse(readFileSync(r,"utf-8"));return createHash("sha256").update(s).digest("hex")!==i.sha256?null:(await this.verifySignature(i.signature,i.libraryVersion,i.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s;try{let o=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(o.status===401||o.status===403)throw new _("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(!o.ok)throw new _(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${o.status})`,"GATEWAY_FETCH_FAILED");s=await o.json();}catch(o){throw o instanceof _?o:new _("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",o)}let i;try{i=await fetch(s.bundleUrl);}catch(o){throw new _("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",o)}if(!i.ok)throw new _(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${i.status})`,"GATEWAY_FETCH_FAILED");let a=await i.arrayBuffer();return {bytes:Buffer.from(a),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:i}=await import('jose'),a=await i(pe,"RS256"),{payload:l}=await s(e,a),o=JSON.parse(new TextDecoder().decode(l));if(o.purpose!=="gateway-bundle")throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(o.libraryVersion!==t)throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(o.sha256!==r)throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof _?s:new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=join(homedir(),".inference-relay","gateway","runtime.cjs"),r=dirname(t),s=createRequire(t),i={exports:{}};compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(i.exports,s,i,t,r);let l=i.exports;if(!l.ClaudeCLIProvider)throw new _("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return l}catch(t){throw t instanceof _?t:new _("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){F.clear();}};var ot="1.4.0",te=class extends A{type="claude-cli";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await B.load(this.licenseKey,ot)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof _?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(ee(),Me));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};var K=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],V=Object.fromEntries(K.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));V["claude-sonnet-4"]="claude-sonnet-4-20250514";V["claude-opus-4"]="claude-opus-4-20250514";V["claude-haiku-4"]="claude-haiku-4-5-20251001";function ge(n){return V[n]||n}function lt(n){return K.find(e=>e.canonical===n)}function re(n){return K.find(e=>e.canonical===n)?.provider}function ct(n){return K.filter(e=>e.provider===n)}var ut="https://api.anthropic.com/v1/messages",dt="2023-06-01",fe={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};function Oe(n){if(fe[n])return fe[n];for(let[e,t]of Object.entries(fe))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var G=class extends A{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||ut,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),l=Date.now()-r;this.recordSuccess();let[o,d]=Oe(a.model||e.model),c=((a.usage?.input_tokens||0)*o+(a.usage?.output_tokens||0)*d)/1e6,f={},h=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let p of h){let E=i.headers.get(p);E&&(f[p]=E);}return {content:a.content||[],model:a.model||e.model,usage:{input_tokens:a.usage?.input_tokens||0,output_tokens:a.usage?.output_tokens||0},stop_reason:a.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:l,fallbackChain:["anthropic-api:success"],headers:Object.keys(f).length>0?f:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder;return (async function*(){let l="",o=false;try{for(;;){let{done:d,value:c}=await i.read();if(d)break;l+=a.decode(c,{stream:!0});let f=l.split(`
+`);l=f.pop()||"";let h=null;for(let p of f)if(p.startsWith("event: "))h=p.slice(7).trim();else if(p.startsWith("data: ")){let E=p.slice(6);if(h&&E)try{let u=JSON.parse(E);u.type==="message_stop"&&(o=!0),yield u;}catch{console.warn("[inference-relay] Skipped malformed SSE data:",E.substring(0,100));}h=null;}else p.trim()===""&&(h=null);}o||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[i,a]=Oe(e.model),l=(s*i+1e3*a)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(l*1e4)/1e4,model:e.model,provider:"anthropic-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){let e={"content-type":"application/json","x-api-key":this.apiKey,"anthropic-version":dt};return this.dangerousDirectBrowserAccess&&(e["anthropic-dangerous-direct-browser-access"]="true"),e}buildRequestBody(e,t){let r={model:ge(e.model),max_tokens:e.max_tokens,messages:e.messages};return t&&(r.stream=true),e.system&&(r.system=e.system),e.temperature!==void 0&&(r.temperature=e.temperature),e.tools&&e.tools.length>0&&(r.tools=e.tools),e.metadata&&(r.metadata=e.metadata),r}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let i=new Error(`ANTHROPIC_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw i.retryAfterMs=s,i}throw t===529?(this.health.status="degraded",this.health.lastError="API overloaded",this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_OVERLOADED: API is temporarily overloaded (529)")):(this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`ANTHROPIC_API_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var pt="https://api.openai.com/v1/chat/completions",De=["gpt-","o1","o3","o4","chatgpt-","text-","davinci","babbage","ft:"],be={"gpt-4.1":[2,8],"gpt-4.1-mini":[.4,1.6],"gpt-4.1-nano":[.1,.4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"o4-mini":[1.1,4.4],o3:[.4,1.6],"o3-mini":[1.1,4.4],o1:[15,60],"o1-mini":[3,12],"gpt-4-turbo":[10,30],"gpt-4":[30,60],"gpt-3.5-turbo":[.5,1.5]};function Ne(n){if(be[n])return be[n];for(let[e,t]of Object.entries(be))if(n.startsWith(e))return t;return [2.5,10]}function mt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function ht(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n){let s=mt(r.content);if(r.role==="assistant"&&Array.isArray(r.content)){let i=r.content.filter(a=>a.type==="tool_use");if(i.length>0){t.push({role:"assistant",content:s||"",tool_calls:i.map(a=>({id:a.id||`call_${Math.random().toString(36).slice(2)}`,type:"function",function:{name:a.name,arguments:typeof a.input=="string"?a.input:JSON.stringify(a.input)}}))});continue}}t.push({role:r.role,content:s});}return t}function Ue(n){switch(n){case "stop":return "end_turn";case "length":return "max_tokens";case "tool_calls":case "function_call":return "tool_use";default:return "end_turn"}}var W=class extends A{type="openai-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||pt;}async probe(){try{let e=this.baseUrl.replace(/\/chat\/completions$/,"/models"),t=await fetch(e,{method:"GET",headers:this.buildHeaders()});return t.status===401?{available:!0,authenticated:!1}:t.ok||t.status===429?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),l=Date.now()-r;this.recordSuccess();let o=a.choices?.[0],d=[];if(o?.message?.content&&d.push({type:"text",text:o.message.content}),o?.message?.tool_calls)for(let m of o.message.tool_calls)d.push({type:"tool_use",id:m.id,name:m.function?.name,input:m.function?.arguments?JSON.parse(m.function.arguments):{}});let c=a.usage?.prompt_tokens||0,f=a.usage?.completion_tokens||0,[h,p]=Ne(a.model||e.model),E=(c*h+f*p)/1e6,u={},g=["x-request-id","x-ratelimit-limit-requests","x-ratelimit-remaining-requests","x-ratelimit-limit-tokens","x-ratelimit-remaining-tokens","x-ratelimit-reset-requests","x-ratelimit-reset-tokens"];for(let m of g){let y=i.headers.get(m);y&&(u[m]=y);}return {content:d,model:a.model||e.model,usage:{input_tokens:c,output_tokens:f},stop_reason:Ue(o?.finish_reason),provider:"openai-api",costUsd:Math.round(E*1e6)/1e6,durationMs:l,fallbackChain:["openai-api:success"],headers:Object.keys(u).length>0?u:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OPENAI_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder,l=e.model;return (async function*(){let o="",d=false,c=false,f,h;yield {type:"message_start",message:{model:l,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:p,value:E}=await i.read();if(p)break;o+=a.decode(E,{stream:!0});let u=o.split(`
+`);o=u.pop()||"";for(let g of u){if(!g.startsWith("data: "))continue;let m=g.slice(6).trim();if(m==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:h||"end_turn"},usage:f||{}},yield {type:"message_stop"},c=!0;return}try{let y=JSON.parse(m);y.usage&&(f={input_tokens:y.usage.prompt_tokens||0,output_tokens:y.usage.completion_tokens||0});let C=y.choices?.[0];if(!C)continue;let b=C.delta;b?.content&&!d&&(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),b?.content&&(yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:b.content}}),C.finish_reason&&(d&&(yield {type:"content_block_stop",index:0}),h=Ue(C.finish_reason));}catch{}}}c||(yield {type:"message_delta",delta:{stop_reason:h||"end_turn"},usage:f||{}},yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[i,a]=Ne(e.model),l=(s*i+1e3*a)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(l*1e4)/1e4,model:e.model,provider:"openai-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){return {"Content-Type":"application/json",Authorization:`Bearer ${this.apiKey}`}}buildRequestBody(e,t){let r=e.model.toLowerCase();if(!De.some(l=>r.startsWith(l)))throw new Error(`OPENAI_MODEL_TRANSLATION_REQUIRED: The OpenAI provider received the model name '${e.model}', which does not match any known OpenAI model prefix (${De.join(", ")}). The relay does NOT auto-translate model names across providers in v1.3.x. Either (a) pass an OpenAI model name like 'gpt-4o-mini' when routing to this provider, or (b) route non-OpenAI models through a provider that understands them (e.g. 'anthropic-api' for Claude models).`);let i=ht(e.messages,e.system),a={model:e.model,max_tokens:e.max_tokens,messages:i};return t&&(a.stream=true,a.stream_options={include_usage:true}),e.temperature!==void 0&&(a.temperature=e.temperature),e.tools&&e.tools.length>0&&(a.tools=e.tools),a}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("OPENAI_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let i=new Error(`OPENAI_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw i.retryAfterMs=s,i}throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`OPENAI_API_ERROR: Request failed with status ${t}`)}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var yt="http://localhost:11434";function gt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function ft(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n)t.push({role:r.role,content:gt(r.content)});return t}var $=class extends A{type="ollama";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.baseUrl=e?.baseUrl?.replace(/\/$/,"")||yt;}async probe(){try{return (await fetch(`${this.baseUrl}/api/tags`,{method:"GET"})).ok?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),l=Date.now()-r;this.recordSuccess();let o=[];a.message?.content&&o.push({type:"text",text:a.message.content});let d=a.prompt_eval_count||0,c=a.eval_count||0;return {content:o,model:a.model||e.model,usage:{input_tokens:d,output_tokens:c},stop_reason:(a.done,"end_turn"),provider:"ollama",costUsd:0,durationMs:l,fallbackChain:["ollama:success"]}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OLLAMA_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder,l=e.model;return (async function*(){let o="",d=false,c=false;yield {type:"message_start",message:{model:l,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:f,value:h}=await i.read();if(f)break;o+=a.decode(h,{stream:!0});let p=o.split(`
+`);o=p.pop()||"";for(let E of p){let u=E.trim();if(u)try{let g=JSON.parse(u);if(g.message?.content&&(d||(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:g.message.content}}),g.done){d&&(yield {type:"content_block_stop",index:0}),yield {type:"message_delta",delta:{stop_reason:"end_turn"},usage:{input_tokens:g.prompt_eval_count||0,output_tokens:g.eval_count||0}},yield {type:"message_stop"},c=!0;return}}catch{}}}c||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0;return {estimatedInputTokens:t+r,estimatedCost:0,model:e.model,provider:"ollama"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildRequestBody(e,t){if(e.tools&&e.tools.length>0)throw new Error(`OLLAMA_TOOLS_NOT_SUPPORTED: The Ollama provider received ${e.tools.length} tool definition(s) in params.tools, but this provider does not implement function calling in v1.3.x. For tool-use workflows route to 'anthropic-api' or 'openai-api' instead. Tracking issue: 1.4.0 ships an Ollama tool-use shim for models that support native function calling (llama3.1, mistral-nemo, etc.).`);let r=ft(e.messages,e.system),s={model:e.model,messages:r,stream:t};return e.temperature!==void 0&&(s.options={temperature:e.temperature}),s}handleErrorResponse(e){let t=e.status;throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),t===404?(this.health.status="unhealthy",this.health.lastError="Model not found",new Error("OLLAMA_MODEL_NOT_FOUND: Requested model is not available locally (404)")):(this.health.status="unhealthy",this.health.lastError=`Request failed with status ${t}`,new Error(`OLLAMA_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var Y=class extends EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};var Fe=new Set(["anthropic-api","openai-api","ollama"]),q=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};function Be(n){return n.messages.some(e=>typeof e.content=="string"?false:Array.isArray(e.content)&&e.content.some(t=>t.type==="image"||t.type==="image_url"))}function _e(n){if(!(n instanceof Error))return "unknown";let e=n.message.toLowerCase(),t=n.name.toLowerCase();return e.startsWith("openai_model_translation_required:")||e.startsWith("ollama_tools_not_supported:")?"provider_incompatible":e.includes("rate limit")||e.includes("429")||t.includes("ratelimit")?"rate_limited":e.includes("auth")||e.includes("401")||e.includes("403")||e.includes("unauthorized")||e.includes("forbidden")?"auth_failed":e.includes("timeout")||e.includes("timed out")||e.includes("aborted")||t==="aborterror"||t==="timeouterror"?"timeout":"unknown"}function He(n){return [...n].sort((e,t)=>e.priority-t.priority)}var j=class n{healthMap=new Map;engineEvents=new Y;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=He(t),i=[],a=Be(e),l=randomUUID(),o=Date.now();for(let d of s){let c=d.type;if(this.getHealthStatus(c)==="unhealthy"){i.push(`${c}:skipped:unhealthy`);continue}if(a&&!d.supportsMultimodal){i.push(`${c}:skipped:no_multimodal`);continue}let h=re(e.model);if(h&&h!==c&&Fe.has(c)){i.push(`${c}:skipped:wrong_provider`);continue}try{let p=await d.execute(e,{signal:r?.signal});return this.markHealthy(c),i.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:l,id:randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:p.usage.input_tokens,outputTokens:p.usage.output_tokens,costUsd:p.costUsd,userCostUsd:p.userCostUsd??0}],provider:c,model:p.model,inputTokens:p.usage.input_tokens,outputTokens:p.usage.output_tokens,costUsd:p.costUsd,userCostUsd:p.userCostUsd,durationMs:p.durationMs??Date.now()-o,stopReason:p.stop_reason,fallbackChain:i,rateLimitHit:!1,securityState:r?.licenseState}),{...p,fallbackChain:i}}catch(p){if(p?.name==="AbortError"||r?.signal?.aborted)throw p;switch(_e(p)){case "rate_limited":d.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),i.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),i.push(`${c}:failed:auth_failed`);break;case "timeout":i.push(`${c}:failed:timeout`);break;case "provider_incompatible":i.push(`${c}:incompatible`);break;default:i.push(`${c}:failed:${p.message?.slice(0,40)||"unknown"}`);break}}}throw new q(i)}async stream(e,t,r){let s=randomUUID(),i=He(t),a=Be(e),l=Date.now(),o=[],d=null,c=null;for(let m of i){let y=m.type;if(this.getHealthStatus(y)==="unhealthy"){o.push(`${y}:skipped:unhealthy`);continue}if(a&&!m.supportsMultimodal){o.push(`${y}:skipped:no_multimodal`);continue}if(!m.supportsStreaming){o.push(`${y}:skipped:no_streaming`);continue}let C=re(e.model);if(C&&C!==y&&Fe.has(y)){o.push(`${y}:skipped:wrong_provider`);continue}try{c=await m.streamRaw(e,{signal:r?.signal}),d=m,o.push(`${y}:success`);break}catch(b){if(b?.name==="AbortError"||r?.signal?.aborted)throw b;switch(_e(b)){case "rate_limited":m.reportRateLimit({provider:y,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(y,"rate_limited"),o.push(`${y}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(y,"auth_failed"),o.push(`${y}:failed:auth_failed`);break;case "timeout":o.push(`${y}:failed:timeout`);break;case "provider_incompatible":o.push(`${y}:incompatible`);break;default:o.push(`${y}:failed:${b.message?.slice(0,40)||"unknown"}`);break}}}if(!d||!c)throw new q(o);let h=this,p=r?.maxMidStreamFailovers??2,E=new Set([d]);async function*u(){let m=d,y=c[Symbol.asyncIterator](),C={role:"assistant",content:[]},b={segments:[]},I=Ke(m.type);b.segments.push(I);let Re=0,z=null,X=false,ue=p,D="",N="";for(h.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:m.type});;)try{for(;;){let R=await y.next();if(R.done)break;let T=R.value,k=vt(T,Re,C,X,z);if(k.action==="suppress"){"coalesceMode"in k&&(z=k.coalesceMode),k.clearedJustFailedOver&&(X=!1);continue}k.coalesceModeCleared&&(z=null),k.clearedJustFailedOver&&(X=!1);let v=k.event;if(Et(C,v),v.type==="message_start"){if(b.segments.length>1)continue;D=v.message?.model||D,v.message?.usage?.input_tokens!=null&&(I.inputTokens=v.message.usage.input_tokens);}if(v.type==="message_delta"&&(v.delta?.stop_reason&&(N=v.delta.stop_reason),v.usage)){_t(I,m.type,v.usage),yield {...v,usage:void 0};continue}if(v.type==="message_stop")break;yield v;}if(C.content.length===0&&I.outputTokens===0)throw new Error("stream ended without producing any content");h.markHealthy(m.type),h.emitEngineEvent({kind:"segment_complete",callId:s,id:randomUUID(),ts:Date.now(),provider:m.type,usage:{...I}});let S=Ve(b.segments);yield {type:"message_delta",delta:{stop_reason:N||"end_turn"},usage:S},yield {type:"message_stop"},h.emitEngineEvent({kind:"stream_complete",callId:s,id:randomUUID(),ts:Date.now(),stitched:b.segments.length>1,segments:b.segments.map(R=>({...R})),provider:b.segments.map(R=>R.provider).join("+"),model:D,inputTokens:S.input_tokens,outputTokens:S.output_tokens,costUsd:b.segments.reduce((R,T)=>R+T.costUsd,0),userCostUsd:b.segments.reduce((R,T)=>R+T.userCostUsd,0),durationMs:Date.now()-l,stopReason:N||"end_turn",fallbackChain:o,rateLimitHit:!1,securityState:r?.licenseState});return}catch(S){if(S?.name==="AbortError"||r?.signal?.aborted)throw S;let R=_e(S);try{await y.return?.();}catch{}if(h.markUnhealthy(m.type,String(S)),h.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:randomUUID(),ts:Date.now(),from:m.type,bytesBuffered:Ge(C),errorClass:R,retriesLeft:ue}),o.push(`${m.type}:partial(${Ge(C)}b)`),ue<=0)throw h.emitTerminalExhaustion(s,b,D,N,o,l,r?.licenseState),S;let T=Ct(C.content),k=[...e.messages,{role:"assistant",content:T}],v=null,Q=null,Ae=m;for(;;){let w=h.nextHealthyProvider(i,Ae,a,E);if(!w)break;Ae=w,E.add(w);let qe={...e,messages:k,...w.type!=="anthropic-api"&&w.type!=="claude-cli"?{system:Rt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{Q=await w.streamRaw(qe,{signal:r?.signal}),v=w;break}catch(Se){h.markUnhealthy(w.type,String(Se)),o.push(`${w.type}:failed_connect:${String(Se).slice(0,40)}`);}}if(!v||!Q)throw h.emitTerminalExhaustion(s,b,D,N,o,l,r?.licenseState),S;o.push(`${v.type}:stitched`),h.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:randomUUID(),ts:Date.now(),from:m.type,to:v.type}),Re=At(C)+1,ue--,m=v,y=Q[Symbol.asyncIterator](),X=true,z=null,I=Ke(m.type),b.segments.push(I),h.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:m.type});}}let g=u();return r?.licenseState==="SEC_DEGRADED"&&(g=St(g)),L(g)}nextHealthyProvider(e,t,r,s){let i=e.indexOf(t);if(i===-1)return null;for(let a=i+1;a<e.length;a++){let l=e[a];if(!s.has(l)&&!(r&&!l.supportsMultimodal)&&l.supportsStreaming)return l}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,i,a,l){let o=Ve(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(d=>({...d})),provider:t.segments.map(d=>d.provider).join("+"),model:r,inputTokens:o.input_tokens,outputTokens:o.output_tokens,costUsd:t.segments.reduce((d,c)=>d+c.costUsd,0),userCostUsd:t.segments.reduce((d,c)=>d+c.userCostUsd,0),durationMs:Date.now()-a,stopReason:s||"error",fallbackChain:i,rateLimitHit:false,securityState:l});}};function Ke(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function _t(n,e,t){typeof t.input_tokens=="number"&&(n.inputTokens=t.input_tokens),typeof t.output_tokens=="number"&&(e==="anthropic-api"||e==="claude-cli"?n.outputTokens+=t.output_tokens:n.outputTokens=t.output_tokens);}function Ve(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function Et(n,e){if(e.type==="content_block_start"){let t=e.content_block;t?.type==="text"?n.content.push({type:"text",text:t.text||"",open:true}):t?.type==="tool_use"&&n.content.push({type:"tool_use",id:t.id,name:t.name,input:{},_partialJson:"",open:true});}else if(e.type==="content_block_delta"){let t=n.content[n.content.length-1];if(!t)return;t.type==="text"&&e.delta?.type==="text_delta"?t.text+=e.delta.text||"":t.type==="tool_use"&&e.delta?.type==="input_json_delta"&&(t._partialJson=(t._partialJson||"")+(e.delta.partial_json||""));}else if(e.type==="content_block_stop"){for(let t=n.content.length-1;t>=0;t--)if(n.content[t].open){if(n.content[t].open=false,n.content[t].type==="tool_use"){let r=n.content[t];try{r.input=JSON.parse(r._partialJson||"{}");}catch{}delete r._partialJson;}break}}}function vt(n,e,t,r,s){if(s&&n.type==="content_block_delta"&&n.index===s.fromIndex)return {action:"yield",event:{...n,index:s.toIndex}};if(s&&n.type==="content_block_stop"&&n.index===s.fromIndex)return {action:"suppress",coalesceMode:null,coalesceModeCleared:true};if(r&&n.type==="content_block_start"&&n.content_block?.type==="text"){let a=t.content[t.content.length-1];if(a?.type==="text"&&a.open===true)return {action:"suppress",coalesceMode:{fromIndex:n.index+e,toIndex:t.content.length-1},clearedJustFailedOver:true}}let i=r&&n.type==="content_block_start";return typeof n.index=="number"&&e>0?{action:"yield",event:{...n,index:n.index+e},clearedJustFailedOver:i}:{action:"yield",event:n,clearedJustFailedOver:i}}function Ct(n){let e=[];for(let t of n)t.type==="text"?t.text&&e.push({type:"text",text:t.text}):t.type==="tool_use"&&!t.open&&t.input&&Object.keys(t.input).length>0&&e.push({type:"tool_use",id:t.id,name:t.name,input:t.input});return e}function Rt(n,e){return n?`${n}
+
+${e}`:e}function Ge(n){let e=0;for(let t of n.content)t.type==="text"?e+=t.text.length:t.type==="tool_use"&&(e+=(t._partialJson||"").length);return e}function At(n){return Math.max(0,n.content.length-1)}async function*St(n){for await(let e of n)await new Promise(t=>setTimeout(t,50+Math.random()*50)),yield e;}ee();function ne(n,e){if(e==="prefer-user"){let s=n.filter(a=>a.type==="claude-cli"),i=n.filter(a=>a.type!=="claude-cli");return [...s,...i]}if(e==="prefer-app"){let s=n.filter(a=>a.type==="claude-cli");return [...n.filter(a=>a.type!=="claude-cli"),...s]}let t=n.filter(s=>s.type===e),r=n.filter(s=>s.type!==e);return [...t,...r]}function se(n){return (e,t)=>n.maxCostPerCall!==void 0&&t.estimatedCost>n.maxCostPerCall?"prefer-user":n.preferLocal?"ollama":"prefer-user"}var Ee={"claude-opus-4":[15,75],"claude-sonnet-4":[3,15],"claude-haiku-4":[.8,4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"gpt-4-turbo":[10,30],"gpt-3.5-turbo":[.5,1.5],llama:[0,0],mistral:[0,0],codellama:[0,0]};function ie(n){if(Ee[n])return Ee[n];for(let[e,t]of Object.entries(Ee))if(n.startsWith(e))return t;return [3,15]}function ae(n){let e=0;for(let l of n.messages)if(typeof l.content=="string")e+=l.content.length;else if(Array.isArray(l.content))for(let o of l.content)o.type==="text"&&o.text&&(e+=o.text.length);n.system&&(e+=n.system.length);let t=Math.ceil(e/4),r=Math.min(n.max_tokens||1024,4096),[s,i]=ie(n.model),a=(t*s+r*i)/1e6;return {estimatedInputTokens:t,estimatedCost:Math.round(a*1e4)/1e4,model:n.model,provider:""}}var O=class{endpoint;enabled;keyId;lineageBuffer=new Map;unsubscribers=[];constructor(e){this.endpoint=e.endpoint||"https://api.inference-relay.com/v1/events",this.enabled=e.enabled??true,this.keyId=e.keyId||"";}subscribeToEngine(e){let t=s=>this.bufferLineageEvent(s),r=s=>{s.kind==="stream_complete"&&this.flushLineageForCall(s);};e.engineEvents.onTyped("segment_start",t),e.engineEvents.onTyped("segment_complete",t),e.engineEvents.onTyped("mid_stream_failover_attempt",t),e.engineEvents.onTyped("mid_stream_failover_success",t),e.engineEvents.onTyped("stream_complete",r),this.unsubscribers.push(()=>{e.engineEvents.offTyped("segment_start",t),e.engineEvents.offTyped("segment_complete",t),e.engineEvents.offTyped("mid_stream_failover_attempt",t),e.engineEvents.offTyped("mid_stream_failover_success",t),e.engineEvents.offTyped("stream_complete",r);});}unsubscribe(){for(let e of this.unsubscribers)e();this.unsubscribers=[],this.lineageBuffer.clear();}bufferLineageEvent(e){let t=this.lineageBuffer.get(e.callId)??[];t.push(e),this.lineageBuffer.set(e.callId,t);}flushLineageForCall(e){let t=this.lineageBuffer.get(e.callId)??[];this.lineageBuffer.delete(e.callId);let r={timestamp:e.ts,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,userCostUsd:e.userCostUsd,durationMs:e.durationMs,stopReason:e.stopReason,fallbackChain:e.fallbackChain,rateLimitHit:e.rateLimitHit,securityState:e.securityState,stitched:e.stitched,segments:e.segments,lineage:t.map(s=>s.id),promptContent:false,completionContent:false};this.report(r);}report(e){!this.enabled||!this.keyId||fetch(this.endpoint,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({keyId:this.keyId,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,...e.userCostUsd!==void 0?{userCostUsd:e.userCostUsd}:{},durationMs:e.durationMs,fallback:e.fallbackChain.length>1,timestamp:e.timestamp,...e.stitched?{stitched:true,segmentCount:e.segments?.length??0}:{},...e.lineage?{lineageIds:e.lineage}:{}})}).catch(()=>{});}};var Pt="https://api.inference-relay.com",kt="/v1/validate",wt=1440*60*1e3,It=10080*60*1e3,Tt=`-----BEGIN PUBLIC KEY-----
+MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
+EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
+t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
+kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
+7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
+Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
+UQIDAQAB
+-----END PUBLIC KEY-----`,x=new Map;async function ve(n,e){if(!n)return {valid:false,error:"License key required. Get one at inference-relay.com"};let t=x.get(n);if(t&&Date.now()-t.cachedAt<t.ttl)return t.result;let r=new AbortController,i=!x.has(n)?3e3:1e3,a=setTimeout(()=>r.abort(),i),l;try{let d=(e||Pt)+kt;l=await(await fetch(d,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({key:n}),signal:r.signal})).json();}catch{return clearTimeout(a),t&&Date.now()-t.cachedAt<It?t.result:{valid:false,error:"License validation failed (network error). Visit inference-relay.com"}}clearTimeout(a);let o;if(l.jws)try{let{compactVerify:d,importSPKI:c}=await import('jose'),f=await c(Tt,"RS256"),{payload:h}=await d(l.jws,f);o=JSON.parse(new TextDecoder().decode(h));}catch{return x.delete(n),{valid:false,error:"License validation signature invalid. Possible MITM \u2014 refusing to start."}}else if(o=l,o.valid)return x.delete(n),{valid:false,error:"License validation response not signed. Possible MITM."};return x.set(n,{result:o,cachedAt:Date.now(),ttl:o.valid?wt:6e4}),o}var Ce=class n{providers=[];config;fallbackEngine;telemetryReporter;routingFn=null;license=null;licensePromise=null;licenseValidatedAt=0;callsSinceValidation=0;manifestDecryptor=new M;capProbeTimer=null;capProbeTicksRemaining=0;messages;constructor(e){if(this.config=e,this.fallbackEngine=new j,this.telemetryReporter=new O({endpoint:e.endpoint?`${e.endpoint}/v1/events`:void 0,enabled:e.telemetry!==false,keyId:e.licenseKey}),this.telemetryReporter.subscribeToEngine(this.fallbackEngine),e.routing&&(this.routingFn=se(e.routing)),e.providers)for(let s of e.providers){let i=typeof s=="string"?{type:s}:s,a=this.createProvider(i);a&&(i.priority!==void 0&&(a.priority=i.priority),this.providers.push(a));}let t=100,r=3600*1e3;this.messages={create:async(s,i)=>{let a=this.license&&(this.callsSinceValidation>=t||Date.now()-this.licenseValidatedAt>=r);if(a&&(this.license=null),!this.license){this.licensePromise||(this.licensePromise=ve(this.config.licenseKey,this.config.endpoint));try{this.license=await this.licensePromise;}finally{this.licensePromise=null;}if(this.licenseValidatedAt=Date.now(),this.callsSinceValidation=0,this.license.valid){if(this.license.capExceeded)throw this.startCapProbe(),new U({tier:this.license.tier||"solo",baseCap:this.license.baseCap??this.license.usageCap??null,topUps:this.license.topUpGrants??0,effectiveCap:this.license.usageCap??null,usedThisCycle:this.license.usageThisMonth??0});a||console.log(`[inference-relay] LICENSE AUTHORIZED [${(this.license.tier||"unknown").toUpperCase()} TIER] \u2014 READY.`);}else throw new Error(`[inference-relay] ${this.license.error||"License key invalid or expired. Visit inference-relay.com"}`)}return this.callsSinceValidation++,s.stream?this.streamMessage(s,i):this.executeMessage(s,i)}};}startCapProbe(){if(this.capProbeTimer!==null)return;let e=15e3,t=20,r=4;this.capProbeTicksRemaining=t;let s=async()=>{this.capProbeTicksRemaining-=1;let i=Math.ceil(this.capProbeTicksRemaining*e/6e4);this.capProbeTicksRemaining>0&&this.capProbeTicksRemaining%r===0&&console.log(`[inference-relay] cap exceeded \u2014 checking for top-up every 15s (${i}m remaining)`),this.config.licenseKey&&x.delete(this.config.licenseKey);try{let a=await this.probeValidate();if(a.valid&&a.capExceeded===!1){console.log("[inference-relay] cap cleared \u2014 probe exiting, inference resuming"),this.license=a,this.licenseValidatedAt=Date.now(),this.callsSinceValidation=0,this.stopCapProbe();return}}catch{}this.capProbeTicksRemaining<=0&&(console.log("[inference-relay] cap probe exhausted (5 min) \u2014 stopping"),this.stopCapProbe());};this.capProbeTimer=setInterval(s,e),typeof this.capProbeTimer.unref=="function"&&this.capProbeTimer.unref();}async probeValidate(){return ve(this.config.licenseKey,this.config.endpoint)}stopCapProbe(){this.capProbeTimer!==null&&(clearInterval(this.capProbeTimer),this.capProbeTimer=null),this.capProbeTicksRemaining=0;}static async autoDetect(e){let t=[],r=await H();return r.available&&r.authenticated&&t.push({type:"claude-cli",priority:1}),process.env.ANTHROPIC_API_KEY&&t.push({type:"anthropic-api",priority:2,apiKey:process.env.ANTHROPIC_API_KEY}),new n({licenseKey:e?.licenseKey||process.env.IR_LICENSE_KEY||"",providers:t,fallback:e?.fallback??true,...e})}dispose(){this.providers=[],this.config.licenseKey&&x.delete(this.config.licenseKey),this.license=null,this.licensePromise=null,this.callsSinceValidation=0,this.stopCapProbe(),this.telemetryReporter.unsubscribe(),this.manifestDecryptor.clearCache(),Pe();}resolveProviders(e){if(!this.routingFn)return this.providers;let t=ae(e),r=this.routingFn(e,t);return ne(this.providers,r)}async executeMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.execute(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState()})}async streamMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.stream(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState()})}createProvider(e){switch(e.type){case "claude-cli":return new te(this.config.licenseKey);case "anthropic-api":return e.apiKey?new G({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "openai-api":return e.apiKey?new W({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "ollama":return new $({baseUrl:e.baseUrl});default:return null}}};ee();var xt="https://api.anthropic.com/v1/messages",Lt="2023-06-01";async function We(n){try{let e=await fetch(xt,{method:"POST",headers:{"content-type":"application/json","x-api-key":n,"anthropic-version":Lt},body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{valid:!1,error:"Invalid API key (401)"}:e.ok?{valid:!0}:e.status===429?{valid:!0,error:"Rate limited (429)"}:e.status===529?{valid:!0,error:"API overloaded (529)"}:{valid:!1,error:`Unexpected status ${e.status}`}}catch(e){return {valid:false,error:`Network error: ${e.message}`}}}var oe=class{filePath;safeStorage=null;memoryFallback=new Map;constructor(e){let t=e?.storagePath||join(homedir(),".inference-relay");existsSync(t)||mkdirSync(t,{recursive:true}),this.filePath=join(t,"credentials.enc");try{let r=Je("electron");r?.safeStorage?.isEncryptionAvailable?.()&&(this.safeStorage=r.safeStorage);}catch{}}readDisk(){if(!existsSync(this.filePath))return {};try{let e=readFileSync(this.filePath,"utf-8");return JSON.parse(e)}catch{return {}}}writeDisk(e){writeFileSync(this.filePath,JSON.stringify(e,null,2),{mode:384});}async get(e){if(!this.safeStorage)return this.memoryFallback.get(e)??null;let r=this.readDisk()[e];if(!r)return null;try{let s=Buffer.from(r.data,"base64");return this.safeStorage.decryptString(s)}catch{return null}}async set(e,t){if(!this.safeStorage){this.memoryFallback.set(e,t);return}let r=this.readDisk(),s=this.safeStorage.encryptString(t);r[e]={data:s.toString("base64")},this.writeDisk(r);}async delete(e){if(!this.safeStorage){this.memoryFallback.delete(e);return}let t=this.readDisk();e in t&&(delete t[e],this.writeDisk(t));}async has(e){if(!this.safeStorage)return this.memoryFallback.has(e);let t=this.readDisk();return e in t}};var J="AES-GCM";var le=class{dbName="inference-relay-credentials";storeName="keys";cryptoKey=null;passphrase;initPromise=null;constructor(e){this.passphrase=e?.passphrase,e?.dbName&&(this.dbName=e.dbName);}async ensureReady(){if(!this.cryptoKey)return this.initPromise?this.initPromise:(this.initPromise=this.init(),this.initPromise)}async init(){if(await this.openDB(),this.passphrase)this.cryptoKey=await this.deriveKey(this.passphrase,this.getOrCreateSessionSalt());else {let e=typeof sessionStorage<"u"?sessionStorage.getItem("ir-cred-key"):null;if(e){let t=Uint8Array.from(atob(e),r=>r.charCodeAt(0));this.cryptoKey=await crypto.subtle.importKey("raw",t,J,false,["encrypt","decrypt"]);}else {this.cryptoKey=await crypto.subtle.generateKey({name:J,length:256},true,["encrypt","decrypt"]);let t=await crypto.subtle.exportKey("raw",this.cryptoKey),r=btoa(String.fromCharCode(...new Uint8Array(t)));typeof sessionStorage<"u"&&sessionStorage.setItem("ir-cred-key",r);}}}openDB(){return new Promise((e,t)=>{let r=indexedDB.open(this.dbName,1);r.onupgradeneeded=()=>{let s=r.result;s.objectStoreNames.contains(this.storeName)||s.createObjectStore(this.storeName);},r.onsuccess=()=>e(r.result),r.onerror=()=>t(r.error);})}async idbGet(e){let t=await this.openDB();return new Promise((r,s)=>{let l=t.transaction(this.storeName,"readonly").objectStore(this.storeName).get(e);l.onsuccess=()=>r(l.result),l.onerror=()=>s(l.error);})}async idbPut(e,t){let r=await this.openDB();return new Promise((s,i)=>{let o=r.transaction(this.storeName,"readwrite").objectStore(this.storeName).put(t,e);o.onsuccess=()=>s(),o.onerror=()=>i(o.error);})}async idbDelete(e){let t=await this.openDB();return new Promise((r,s)=>{let l=t.transaction(this.storeName,"readwrite").objectStore(this.storeName).delete(e);l.onsuccess=()=>r(),l.onerror=()=>s(l.error);})}async idbHas(e){let t=await this.openDB();return new Promise((r,s)=>{let l=t.transaction(this.storeName,"readonly").objectStore(this.storeName).count(e);l.onsuccess=()=>r(l.result>0),l.onerror=()=>s(l.error);})}getOrCreateSessionSalt(){let e="ir-cred-salt";if(typeof sessionStorage<"u"){let r=sessionStorage.getItem(e);if(r)return Uint8Array.from(atob(r),s=>s.charCodeAt(0))}let t=crypto.getRandomValues(new Uint8Array(16));return typeof sessionStorage<"u"&&sessionStorage.setItem(e,btoa(String.fromCharCode(...t))),t}async deriveKey(e,t){let r=new TextEncoder,s=await crypto.subtle.importKey("raw",r.encode(e),"PBKDF2",false,["deriveKey"]);return crypto.subtle.deriveKey({name:"PBKDF2",salt:t.buffer,iterations:1e5,hash:"SHA-256"},s,{name:J,length:256},false,["encrypt","decrypt"])}async encrypt(e){let t=crypto.getRandomValues(new Uint8Array(12)),r=this.getOrCreateSessionSalt(),s=new TextEncoder().encode(e),i=await crypto.subtle.encrypt({name:J,iv:t},this.cryptoKey,s);return {iv:btoa(String.fromCharCode(...t)),ciphertext:btoa(String.fromCharCode(...new Uint8Array(i))),salt:btoa(String.fromCharCode(...r))}}async decrypt(e){let t=Uint8Array.from(atob(e.iv),i=>i.charCodeAt(0)),r=Uint8Array.from(atob(e.ciphertext),i=>i.charCodeAt(0)),s=await crypto.subtle.decrypt({name:J,iv:t},this.cryptoKey,r);return new TextDecoder().decode(s)}async get(e){await this.ensureReady();let t=await this.idbGet(e);if(!t)return null;try{return await this.decrypt(t)}catch{return null}}async set(e,t){await this.ensureReady();let r=await this.encrypt(t);await this.idbPut(e,r);}async delete(e){await this.ensureReady(),await this.idbDelete(e);}async has(e){return await this.ensureReady(),this.idbHas(e)}};var ce=class{store=new Map;async get(e){return this.store.get(e)??null}async set(e,t){this.store.set(e,t);}async delete(e){this.store.delete(e);}async has(e){return this.store.has(e)}};
+export{G as AnthropicAPIProvider,A as BaseProvider,oe as ElectronCredentialStore,Y as EngineEventEmitter,j as FallbackEngine,B as GatewayLoader,_ as GatewayLoaderError,le as IndexedDBCredentialStore,Ce as InferenceRelay,V as MODEL_ALIASES,K as MODEL_REGISTRY,M as ManifestDecryptor,ce as MemoryCredentialStore,$ as OllamaProvider,W as OpenAIAPIProvider,U as RelayCapExceededError,q as RelayError,O as TelemetryReporter,ne as applyRouting,L as createRelayStream,se as createRoutingFunction,H as detectCLI,ae as estimateCost,lt as findModel,Ze as getLastKnownGoodManifest,ie as getModelPricing,et as getValue,ct as modelsForProvider,ge as normalizeModel,re as providerForModel,We as validateAnthropicKey,ve as validateLicense};

package/dist/instruction/index.cjs

@@ -0,0 +1 @@
+'use strict';

package/dist/instruction/index.d.cts

@@ -0,0 +1,2 @@
+
+export {  }

package/dist/instruction/index.d.ts

@@ -0,0 +1,2 @@
+
+export {  }

package/dist/pool/index.cjs

@@ -0,0 +1 @@
+'use strict';

package/dist/pool/index.d.cts

@@ -0,0 +1,2 @@
+
+export {  }

package/dist/pool/index.d.ts

@@ -0,0 +1,2 @@
+
+export {  }

package/dist/stream/index.cjs

@@ -0,0 +1,2 @@
+'use strict';/*! inference-relay (c) 2026 L2B II LLC. Patent pending. Proprietary. See LICENSE. */
+function b(p,r){let s=new Map,a=null,o=null,f=new Promise((e,t)=>{a=e,o=t;});f.catch(()=>{});let l=[],n=null,c="",m="";function i(e,...t){let y=s.get(e);y&&y.forEach(g=>g(...t));}async function*v(){try{for await(let e of p)if(yield e,e.type==="content_block_start")l[e.index]=e.content_block;else if(e.type==="content_block_delta"){let t=l[e.index];t&&e.delta?.type==="text_delta"&&(t.text=(t.text||"")+e.delta.text,i("text",e.delta.text));}else if(e.type==="message_start")c=e.message?.model||"",n=e.message?.usage;else if(e.type==="message_delta")m=e.delta?.stop_reason||"",e.usage&&(n={...n,...e.usage});else if(e.type==="message_stop"){let t={role:"assistant",model:c,content:l.filter(Boolean),usage:n,stop_reason:m};i("message",t),a&&(a(t),a=null),r&&r(t);}if(a){let e={role:"assistant",model:c,content:l.filter(Boolean),usage:n,stop_reason:m||"end_turn"};i("message",e),a(e),a=null,r&&r(e);}}catch(e){throw i("error",e),o&&(o(e),o=null),e}}let _=v(),u=false,d={[Symbol.asyncIterator](){if(u)throw new Error("[inference-relay] RelayStream can only be iterated once");return u=true,_},on(e,t){return s.has(e)||s.set(e,[]),s.get(e).push(t),d},finalMessage(){let e,t=new Promise((y,g)=>{e=setTimeout(()=>g(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([f,t]).finally(()=>clearTimeout(e))}};return d}exports.createRelayStream=b;

package/dist/stream/index.d.cts

@@ -0,0 +1,12 @@
+interface RelayStream<T = any> extends AsyncIterable<T> {
+    on(event: 'text', callback: (text: string) => void): RelayStream<T>;
+    on(event: 'message', callback: (message: any) => void): RelayStream<T>;
+    on(event: 'error', callback: (error: Error) => void): RelayStream<T>;
+    on(event: 'fallback', callback: (from: string, to: string, reason: string) => void): RelayStream<T>;
+    on(event: string, callback: (...args: any[]) => void): RelayStream<T>;
+    /** Resolves when the stream completes. The stream MUST be iterated for this to resolve. */
+    finalMessage(): Promise<any>;
+}
+declare function createRelayStream(eventSource: AsyncIterable<any>, onComplete?: (message: any) => void): RelayStream;
+
+export { type RelayStream, createRelayStream };

package/dist/stream/index.d.ts

@@ -0,0 +1,12 @@
+interface RelayStream<T = any> extends AsyncIterable<T> {
+    on(event: 'text', callback: (text: string) => void): RelayStream<T>;
+    on(event: 'message', callback: (message: any) => void): RelayStream<T>;
+    on(event: 'error', callback: (error: Error) => void): RelayStream<T>;
+    on(event: 'fallback', callback: (from: string, to: string, reason: string) => void): RelayStream<T>;
+    on(event: string, callback: (...args: any[]) => void): RelayStream<T>;
+    /** Resolves when the stream completes. The stream MUST be iterated for this to resolve. */
+    finalMessage(): Promise<any>;
+}
+declare function createRelayStream(eventSource: AsyncIterable<any>, onComplete?: (message: any) => void): RelayStream;
+
+export { type RelayStream, createRelayStream };

package/dist/stream/index.js

@@ -0,0 +1,2 @@
+/*! inference-relay (c) 2026 L2B II LLC. Patent pending. Proprietary. See LICENSE. */
+function b(p,r){let s=new Map,a=null,o=null,f=new Promise((e,t)=>{a=e,o=t;});f.catch(()=>{});let l=[],n=null,c="",m="";function i(e,...t){let y=s.get(e);y&&y.forEach(g=>g(...t));}async function*v(){try{for await(let e of p)if(yield e,e.type==="content_block_start")l[e.index]=e.content_block;else if(e.type==="content_block_delta"){let t=l[e.index];t&&e.delta?.type==="text_delta"&&(t.text=(t.text||"")+e.delta.text,i("text",e.delta.text));}else if(e.type==="message_start")c=e.message?.model||"",n=e.message?.usage;else if(e.type==="message_delta")m=e.delta?.stop_reason||"",e.usage&&(n={...n,...e.usage});else if(e.type==="message_stop"){let t={role:"assistant",model:c,content:l.filter(Boolean),usage:n,stop_reason:m};i("message",t),a&&(a(t),a=null),r&&r(t);}if(a){let e={role:"assistant",model:c,content:l.filter(Boolean),usage:n,stop_reason:m||"end_turn"};i("message",e),a(e),a=null,r&&r(e);}}catch(e){throw i("error",e),o&&(o(e),o=null),e}}let _=v(),u=false,d={[Symbol.asyncIterator](){if(u)throw new Error("[inference-relay] RelayStream can only be iterated once");return u=true,_},on(e,t){return s.has(e)||s.set(e,[]),s.get(e).push(t),d},finalMessage(){let e,t=new Promise((y,g)=>{e=setTimeout(()=>g(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([f,t]).finally(()=>clearTimeout(e))}};return d}export{b as createRelayStream};

package/package.json

@@ -0,0 +1,83 @@
+{
+  "name": "inference-relay",
+  "version": "1.4.0",
+  "private": false,
+  "description": "Route AI inference to the user's existing subscription",
+  "type": "module",
+  "main": "./dist/index.cjs",
+  "module": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "bin": {
+    "inference-relay": "./dist/cli/main.js"
+  },
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js",
+      "require": "./dist/index.cjs"
+    },
+    "./auto": {
+      "import": "./dist/auto.js"
+    },
+    "./instruction": {
+      "types": "./dist/instruction/index.d.ts",
+      "import": "./dist/instruction/index.js"
+    },
+    "./analytics": {
+      "types": "./dist/analytics/index.d.ts",
+      "import": "./dist/analytics/index.js"
+    },
+    "./pool": {
+      "types": "./dist/pool/index.d.ts",
+      "import": "./dist/pool/index.js"
+    },
+    "./stream": {
+      "types": "./dist/stream/index.d.ts",
+      "import": "./dist/stream/index.js"
+    }
+  },
+  "sideEffects": [
+    "./dist/auto.js"
+  ],
+  "peerDependencies": {
+    "@anthropic-ai/sdk": ">=0.50.0"
+  },
+  "peerDependenciesMeta": {
+    "@anthropic-ai/sdk": {
+      "optional": true
+    }
+  },
+  "files": [
+    "dist",
+    "README.md",
+    "LICENSE"
+  ],
+  "scripts": {
+    "build": "tsup",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "lint": "eslint src/",
+    "typecheck": "tsc --noEmit",
+    "clean": "rm -rf dist",
+    "prepublishOnly": "echo 'inference-relay@1.4.0 publish in progress (guard temporarily disabled)'"
+  },
+  "devDependencies": {
+    "@types/node": "^22.0.0",
+    "eslint": "^9.0.0",
+    "tsup": "^8.0.0",
+    "vitest": "^2.0.0"
+  },
+  "engines": {
+    "node": ">=18.0.0"
+  },
+  "dependencies": {
+    "jose": "^6.2.2"
+  },
+  "author": "L2B II LLC",
+  "license": "SEE LICENSE IN LICENSE",
+  "homepage": "https://inference-relay.com",
+  "bugs": {
+    "url": "mailto:support@inference-relay.dev"
+  },
+  "keywords": ["ai", "inference", "claude", "openai", "anthropic", "relay", "routing", "streaming", "fallback", "cost-optimization"]
+}